[
https://issues.apache.org/jira/browse/MAHOUT-796?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13093388#comment-13093388
]
Dmitriy Lyubimov commented on MAHOUT-796:
-----------------------------------------
I re-did dense test to construct 20,000x1,000 dense matrix (20 mln non-zero
elements) with random singular vectors
The way i construct input is i generate random singular vector matrices and
orthogonalize them using stable Gramm-Schmidt, multiply one of them (whatever
is shorter) by Sigma and then produce row-wise surrogate input.
For predefined singular values = 10,4,1,(0.1...), n=1000, m=20000, k=3, p=10 i
get stochastic values
--SSVD solver singular values:
svs: 9.998401 3.998322 0.972622 0.100000 0.100000 0.100000 0.100000
0.100000 0.100000 0.100000 0.100000 0.100000 0.100000
so you see if the decay is good then precision loss with 1 pass in my case
doesn't exceed 2.8% in the worst case (3rd value) and the time is quite good.
(same brunch as for mahout-797 in my github).
Keep in mind that this precision loss also includes loss generated during
simulated input construction, it's not all the solver's.
I also got rid of BBt job and fixed problems with sparse input on that branch.
> Modified power iterations in existing SSVD code
> -----------------------------------------------
>
> Key: MAHOUT-796
> URL: https://issues.apache.org/jira/browse/MAHOUT-796
> Project: Mahout
> Issue Type: Improvement
> Components: Math
> Affects Versions: 0.5
> Reporter: Dmitriy Lyubimov
> Assignee: Dmitriy Lyubimov
> Labels: SSVD
> Fix For: 0.6
>
>
> Nathan Halko contacted me and pointed out importance of availability of power
> iterations and their significant effect on accuracy of smaller eigenvalues
> and noise attenuation.
> Essentially, we would like to introduce yet another job parameter, q, that
> governs amount of optional power iterations. The suggestion how to modify the
> algorithm is outlined here :
> https://github.com/dlyubimov/ssvd-lsi/wiki/Power-iterations-scratchpad .
> Note that it is different from original power iterations formula in the paper
> in the sense that additional orthogonalization performed after each
> iteration. Nathan points out that that improves errors in smaller eigenvalues
> a lot (If i interpret it right).
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
